Efficiency of the KLT on Voiced & Unvoiced Speech as a function of segment size

A novel analysis of the performance of the Karhunen Loeve Transform (KLT) for Voiced and Unvoiced Speech as a function of accepted industry standard speech segment sizes is presented. The speech segment sizes were chosen such that the size must be large enough to capture the properties of the process, but not so large that the signal loses its stationary properties. The energy compaction properties of the KLT are measured and presented as Cumulative Distribution Function (CDF) plots for several unique sets of Eigenvectors. Several CDFs are presented for each segment size to represent the results of each unique set of Eigenvectors. CDFs for DCT and Haar transforms are also presented to serve as a baseline for comparison purposes. The CDF plots not only demonstrate the point of diminishing returns for using the KLT over the DCT or Haar transforms, but also indicate a strong case for a KLT based approach for the purposes of speaker recognition.